Parallel Sparse Coding for Seafloor Image Analysis

Sparse coding has been a popular learning model in machine learning field. However, due to the complexity of the learning model, the high computational cost has seriously hindered its application. Toward this purpose, this paper presents a parallel sparse coding method to improve the performance by exploiting the power of acceleration technologies such as Intel MIC and GPU. We use both parallel programming modes, i.e., the Native model and the Offload model, to parallelize the sparse coding on MIC based computer cluster. Extensive experimental results on the AUV data of the southeast coast of Tasmania have shown that sparse coding can be accelerated significantly on MIC and GPU. When using the same number of threads, the Native model and the Offload model achieve very close performance for sparse coding. In addition, Native model demonstrates better performance scalability than the Offload model. On the other side, parallel implementation on GPU shows the best performance.